Traffic Sign Recognition Report

Traffic Sign Recognition
TRAFFIC SIGN RECOGNITION

A
Project Report
Submitted for the partial fulfilment of
B.Tech. Degree
in
COMPUTER SCIENCE & ENGINEERING
by
SRISHTI KASHYAP (1900520139004)
TEHREEM ARSHAD (1900520139005)
VIVEK PANDEY (1900520139006)
Under the supervision of

PROF. VINEET KANSAL
DR. ADITI SHARMA
Department of Computer Science and Engineering

Institute of Engineering and Technology
Dr. APJ Abdul Kalam Technical University, Lucknow, Uttar Pradesh
May, 30
1
Declaration
We believe that this submission is our job and, within our knowledge and beliefs, materials
previously published or created by someone else, or the awarding of a diploma or diploma for the
following reasons: We hereby declare that we do not include the materials accepted for the serious
error of the university or other university unless recognition is given in the text. This project was
not submitted to another institution to require a different degree.
Submitted by: - Date:24/05/2022
(1) Name: Srishti Kashyap
Roll No: 1900520139004
Branch: Information Technology
Signature:
(2) Name: Tehreem Arshad
Roll No: 1900520139005
Signature:
(3) Name: Vivek Pandey
Roll No: 1900520139006
Signature:
2
Certificate
This is to certify that the project report entitled “Traffic Sign Recognition” presented by
Srishti Kashyap, Tehreem Arshad, and Vivek Pandey for the award of Bachelor of
Technology in Information Technology, is a record of work carried out by them under my
supervision and guidance at the Department of Computer Science and Engineering at Institute
of Engineering and Technology, Lucknow. It is also certified that this project has not been
submitted to any other institute for the award of any other degrees to the best of my knowledge.
Prof. Vineet Kansal

Department of computer science and Engineering
Institute of Engineering and Technology, Lucknow
Dr. Aditi Sharma

Department of Computer science and Engineering
Institute of Engineering and Technology, Lucknow
3
Acknowledgment
At the outset, we are indebted and grateful to our Lord who blessed us with innumerable favors
to reach the present academic heights. We thank God for giving us kind parents so that we were
able to gain knowledge through their compassionate sacrifice and tenacious efforts.
We would like to express our sincerest gratitude to Prof. Vineet Kansal and Dr. Aditi Sharma
of the Institute of Engineering and Technology, Lucknow for all the wisdom and guidance that
they had shown us throughout this year. They had always been gracious and respectful in all
stages of our project and without their support, this project would not have been possible.
We are deeply thankful to our Supervisor (Prof. Vineet Kansal) and co-supervisor (Dr. Aditi
Sharma) for their cooperation and for giving us valuable suggestions and guidance on our
project and for making it correct every time we needed. without their guidance, this project
would not have been possible.
Our deep sense of appreciation goes to our parents, sisters, brothers, and friends for their
continuous support, care and blessings.
Last but not least we are also thankful to our various teachers who taught us various subjects
throughout this course which help us to reach this stage.
Project Members:
SRISHTI KASHYAP (1900520139004)
TEHREEM ARSHAD (1900520139005)
VIVEK PANDEY (1900520139006)
4
Table of Contents
PAGE NO
DECLARATION ................................................................................................................... i
CERTIFICATE.......................................................................................................................ii
ACKNOWLEDGMENT ...................................................................................................... iii
ABSTRACT............................................................................................................................iv
LIST OF FIGURES ............................................................................................................... v
1. INTRODUCTION.............................................................................................................1
1.1 MOTIVATION ............................................................................................................. 1
1.2 OBJECTIVE AND SCOPE… ................................................................................... 1-2
1.3 ORGANIZATION… .................................................................................................... 2
2. LITERATURE REVIEW… ........................................................................................... 3

2.1 TRAFFIC SIGNS…......................................................................................................3
2.2 RELATED WORK… ............................................................................................... 4-5
3. METHODOLOGY… ....................................................................................................... 6
3.1 DESIGN AND IMPLEMENTATION ......................................................................... 6
3.1.1 PROPOSED MODEL… ............................................................................... 6-9
3.1.2 DATA AUGMENTATION… .................................................................. 10-11
3.2 REQUIREMENT ANALYSIS .................................................................................. 11
3.3 DATASET… ........................................................................................................ 11-13
3.4 STEPS FOR TRAFFIC SIGN CLASSIFICATION… ............................................... 13
3.4.1 IMAGE PROCESSING .................................................................................. 13
3.4.2 DETECTION .................................................................................................. 14
3.4.3 CLASSIFICATION… .................................................................................... 14
3.5 FLOW CHART… ...................................................................................................... 15
3.6 IMPLEMENTATION ................................................................................................ 15
3.6.1 DESIGNING A CNN MODEL… ............................................................. 15-16
3.6.2 TRAINING A CNN MODEL… ..................................................................... 17
3.7 ACTIVITY DIAGRAM..............................................................................................18
4. RESULT AND DISCUSSION… ............................................................................ 19-25
5. CONCLUSION… .......................................................................................................... 26
5.1 CONCLUSION… ...................................................................................................... 26
5.2 FUTURE WORK… ................................................................................................... 27
REFERENCES..................................................................................................................... 28
5
ABSTRACT
"Traffic sign recognition and classification" plays an important role in our daily lives today.
These days, road signs are everywhere on the road. Despite this, drivers frequently make
mistakes. While driving on the road, it is quite difficult to perceive and detecttraffic signs. This
may happen due to the following conditions; if the weather conditions are not good and the
signboards become blurry, covered by snow, there is heavy rainfall, signboards are of different
sizes, and they may be rotated by certain angles. Drivers may misunderstand traffic signs. This
leads to accidents, fatal accidents, and damage to the vehicle. To solve this problem, this project
introduces a concept called traffic sign recognition. This model is built using CNN to extract
images and classify road signs. There are different types of traffic signs, For Example, Speed
limit, no entry, traffic lights, left or right turns, children crossing, overtaking of large vehicles,
etc. An important part of the Intelligent Transport Systems is the Traffic Sign Recognition
System (TSRS).
Accurate and effective recognition of traffic signs can increase driving safety. Traffic signs
have certain properties that can be used for recognition and classification. For example, color,
shape, and shading are important attributes that help drivers get road information. The colors
used for traffic signs in each country are almost the same, consisting of colors such as red, blue,
and yellow, and fixed shapes such as circles, triangles, and rectangles. The traffic sign
recognition system is implemented using the Convolutional Neural Network. (CNN). The CNN
model for this project will be used for classification. The project uses the Kaggle German
Traffic Sign Recognition Benchmark as a dataset.
6
ACRONYMS
CNN Convolutional Neural Network
ADAS Advanced driver assistance systems
SDC Self-Driving Car
CV Computer Vision
ITS Intelligent Transformation system
TSR Traffic Sign Recognition
GTSRB German Traffic Sign Recognition Benchmark
ROI Region of interest
IJCNN International joint Conference of Neural Network
7
LIST OF FIGURES
Figure No. Figure Name Page No.
2.1 Traffic Sign of GTSRB 4
3.1(a) Training Phase 6
3.1(b) Testing phase 6
3.2 Rectified Linear Unit activation function 7
3.3 Using Dropout with CNN 8
3.4 Flattening Process 9
3.5 Working on CNN Model 9
3.6 Data Augmentation 10
3.7 Results after applying data Augmentation 11
3.8 GTSRB Dataset 12
3.9 Distribution of Training Dataset 13
3.10 Steps for Traffic Sign Recognition 14
3.11 Flowchart of the Overall process 15
3.12 Basic Building Block of Model 16
3.13 Implementation of the CNN Model 17
3.14(a) Training Phase 18
3.14(b) Training Phase 18
3.15 Activity Diagram 19
4.1 Accuracy Graph of the Model 20
4.2 Loss Graph of the Model 21
4.3 Output is detected with class 28 which is 22

children crossing
4.4 Output is detected with class 28 which is road 22
work.
8
4.5. Output is detected with class 1 which is speed 23

limit of 30 km/h.
4.6 Output is detected with class 5 which is speed 23
limit of 80 km/h.
4.7 GUI of model 24
4.8 Upload an Image to Classify 24
4.9 Output is detected with class 38 which is keep 25

right
4.10 Output is detected with class 38 which is Stop 25
4.11 Output is detected with class 38 which is 26

Slippery Road
9
CHAPTER 1
INTRODUCTION
Traffic Sign Recognition is an important technology that is used to detect and classify traffic
signs. There are various traffic signs present as we move on the road such as children crossing,
road work, 20-speed limit, zebra crossing, no entry, U-turn, etc., we need to follow these traffic
signs for driving safely or to move safely on the roads. Whether we talk about self-driving cars
or human driving cars both the safety of the vehicle and the driver is an important aspect for
which to follow the traffic rules are important. A convolution Neural Network is a Deep
Learning Algorithm that helps to implement this concept of Traffic Sign Recognition with the
help of image classification by taking an image as an input and classifying it with an appropriate
label.
1.1 Motivation
Traffic sign detection and classification is an important technology, as it helps drivers in
understanding signs and following traffic rules, as well as contributing to the development of
autonomous driving systems. Traffic signs present on the roads provide proper assistance to
the drivers which helps them to be alert in advance for the speed regulations, road constructions,
slippery conditions, etc. These traffic signs can be categorized as follows, namely warning,
regulatory, direction signs, and information. Road signs are placed on the road in various sizes
and shapes. Some traffic signs are very blurry and small, covered with snow and heavy rain,
and can rotate at a specific angle so that the driver will not recognize or misunderstand them.
Even in this state, we are aiming to recognize traffic signs. Traffic sign classification is a
technique that uses images as input to recognize traffic signs such as "U-turn destination",
"speed limit 40", and "caution". These signs guide the driver. Therefore, traffic sign recognition
is a challenging topic and a valuable topic in traffic engineering research.
1.2 Objective and scope
Traffic signs are placed next to, above, or adjacent to highways, roads, paths, or other routes
to guide and warn the flow of traffic, including cars, bicycles, pedestrians, riders, and other
travelers. , And a device to regulate. As customer safety becomes more and more important,
traffic sign recognition (TSR) has become one of today's research topics for improving driving
safety. In addition, the resulting system is unaffected by human problems such as fatigue and
10
insomnia. This has improved driving safety.
To this end, TSR systems have been developed primarily to reduce the chance of overlooking
important traffic signs on the road. At first glance, the problem seems easy to solve. However,
the human brain performs steps based on all the experience of visual data, so computers cannot
easily perform the same process. Instead of these experiences, a basic understanding of the
characteristics of traffic indicators can be used. Color and shape information makes up the
bulk of these attributes. This knowledge is not enough to distinguish road signs from others,
but you can improve segmentation by combining intelligence and knowledge. Automatic
traffic sign recognition systems need to solve problems such as sign recognition and location
in crowded scenes, sign feature recognition, and character classification. Implementing traffic
sign classification to identify traffic signs more quickly. To locate Traffic Sign Detection, this
detects the traffic sign as itapproaches us and alerts and notifies us. To learn how to recognize
traffic signs from real-timewebcams to detect traffic signals.
1.3 Organization
The report is divided into five chapters. Chapter 1 gives a brief idea about the project. It tells
about the Objective and Motivation of the project. Chapter 2 is the Literature Review which
provides depth analysis of all the Research Paper read during the project to explore the topic
and Documentation that were used in making this project. Chapter 3 is an explanation of the
methodology used in this project which gives an analysis of how the project will be working
when put under various test case scenarios. Chapter 4 shows the experimental result and all the
possible output. Chapter 5 is concluding the project and possible future work.
11
CHAPTER 2
LITERATURE REVIEW
For the development of a model that can recognize traffic signs through images, previous
research work on the way is reviewed to get the knowledge about the work that had already
been done in this field. The last decade has shown growth trends in the development of
Intelligent Transport Systems (ITS), especially ADAS and Self-Driving Vehicles (SDC). In
these systems, the recognition and recognition of traffic signs is one of the difficult tasks faced
by researchers and developers. This issue has been addressed as a problem of recognizing,
recognizing, and classifying objects (traffic signs) using CVs, but there are still challenges.
2.1 Traffic Signs
Understanding every traffic sign is the most important aspect for all road users. The main rules
and regulations of the road are communicated through traffic signboards that are placed on the
sides of the roads using that can be understood easily within seconds. In addition, everyone
who wants to apply for a driver's license must first understand all of the traffic signs to pass the
driving theory test. The development of self-driving vehicles using the concept of traffic sign
recognition for driving safely on the road is very important as it is very useful in recognizing
roads, vehicles, pedestrians, and traffic signs. The basic principle of a driving assistance system
for traffic sign recognition is to recognize the sign, interpret its meaning, and then send the
information to the driver (via the windshield, screen, or smartphone projection) or even more.
Is to send to the vehicle itself. Then perform the action without the need for a human decision.
The 43 classes that have been used in this project are displayed below that show the basic traffic
signs that are used on the roads to provide information to the driver for his safety, to prevent
accidents, deaths, and, damage to the vehicle.
12
Figure 2.1: Traffic Sign of GTSRB
2.2 Related Work
Deep learning is a branch of machine learning that includes a "convolutional neural network"
used in this project to build a model that classifies images with appropriate labels. Deep
learning algorithms mimic the behavior of the human brain, but at a much smaller scale. Image
classification is used to extract features from an image to identify trends in a dataset.
Ying sun, Pingshu Ge and Dequan Liu (2019) offered a system for traffic sign recognition on
account of deep learning that was used to classify circular traffic signs. This method was used
to identify traffic signs by uniting image preprocessing, traffic sign detection, recognition, and
classification. The accuracy of the system proposed is 98.2 percent, according to the test results
[1].
Md Tarequl Islam (2019) proposed a system that focuses to detect and classify traffic signs that
are circular and triangular in shape. The background color of the traffic signs used was either
red or blue. A set of 28 traffic signs was used for the proposed model. The dataset used was
taken from the UK traffic sign image database. Firstly, the image was blurred using gaussian
blur for smoothing of the image i.e., if we talk in terms of image processing the sharp edges of
the image are eliminated and then converted in the HSV (Hue saturation Value) color space.
To eliminate the blue and red colors from the image masking was done. It is a process in which
certain portions of the image are revealed are some are hidden. The accuracy of the system was
moderate and can be improved by trying different neural network structures. To find the
accuracy used the street images of the UK and Bangladesh. Generally, the system was divided
13
into two stages namely detection and classification. But this is not used for real-time it just
classifies the image [2].
Vaibhavi Golgire (2021) proposed a traffic detection and recognition system and a method
for extracting traffic signs from complex images of nature, processing them, and alerting the
driver with voice commands. It is designed to allow drivers to make quick decisions. The
proposed architecture is divided into three parts, the first part is image preprocessing including
quantification of the input file of the dataset, the second part is to determine the input size for
learning purposes, and The last part is scaling. Information on learning steps. The algorithm
used classifies the symbols recognized during the recognition process [3].
Wenhui li, Daihui, and Shangyou Zeng (2019) worked on both a traditional Convolution Neural
Network and an improved Convolution Neural Network and used the German Traffic Sign
Recognition Benchmark (GTSRB) dataset and the Belgium Traffic Sign Dataset. The
comparison of both the Algorithms was done and the results showed that the accuracy of the
model build using the improved CNN was much higher as compared to the traditional CNN.
The improved CNN uses fewer parameters, smaller models, and easier training i.e., this
algorithm was superior to the traditional CNN method in traffic identification.
14
CHAPTER 3
METHODOLOGY
3.1 Design and Implementation
Input Normalize CNN Train CNN Weight
(a) Training phase
Traffic sign
Input Normalize CNN
Detection
CNN
Weight
Figure 3.1: (b) Testing phase

3.1. 1 Proposed Model
The architecture of a “Convolutional Neural Network” consists of an input layer, some

convolutional layers, a pooling layer, a flattening layer, some fully-connected layers, and an
output layer. CNN is a multi-layered network that functions similarly to the human brain.
Multiple neurons make each layer of CNN.
1. Convolution Layer:
This layer is the primary layer withinside the CNN and a main constructing block withinside
the convolution process. It makes use of convolution to apprehend diverse functions in a given
photo, i.e., it's miles used for function extraction with the assist of filters. It scans the complete
pixel grid after which does a dot product. A filter, every now and then referred to as a kernel,
is a function derived from numerous functions in an enter photo that we need to apprehend.
For example, withinside the case of aspect detection, we'd have separate filters for curves,
blur, and sprucing the photo. As we development deeper into the network, greater state-of-
the-art features emerge as visible.
15
2. Pooling Layer:
This layer is used to downsample the features of the image. Reduces the dimensions of large
images while preserving important details. Reducing the size of the image reduces the
computational power and time required to process the image. Two types of pooling are
available. Depending on the situation, you can use maximum or average pooling. Maximum
pooling uses the maximum value in the feature map, and averaging takes the average of all
pixels.
3. Activation Function:
An Activation Function is used in the convolution layer which transforms the given input into
the required output that has a certain range. This layer introduces non-linear properties to the
network so that the model can predict complex patterns in the data as in the case of images,
audio, video, and, sound. It aids in determining which data should be processed further and
which should not. The weighted sum of the inputs is used as the input to the activation function
that produces the output. This step is important because a model without an activation function
is a simple linear regression model with limited learning ability. The model here uses the ReLU
activation function which replaces all the negative values from zero and if the value is positive
then the output is itself the input. This function has a certain range and is defined by the
following equation.
Rectified Liner Unit: F(u) = max (0, u)
Figure 3.2: Rectified Liner Unit Activation Function
4. Dropout Layer:
16
The dropout layer is used to prevent the problem of overfitting and reduces the complexity of
the model. The problem of overfitting occurs when the model tries to fit more data than what
is actually required and tries to capture every detail fed. As a result, it captures noise and
inaccurate data from the dataset hence the accuracy and performance of the model are reduced.
An overfitted model does not function well.
The Dropout layer used with the convolution neural network is a mask that is used to nullify the
contribution of some of the neurons towards the next layer and all the others are left
unmodified.
If the Dropout layer is not present then the learning is influenced by the first batch of the
training samples in a disproportionately higher manner. Hence the presence of dropout layer in
the training of the CNN models is important.
The combination of the convolution layer, pooling layer, and dropout layer is called phase
learning and is repeated multiple times to improve the training.
Figure 3.3 Using dropout with CNN
5. Flattening Layer:
Pooling layers are output in the form of 3D feature maps.The function of the flatteninglayer is
to transform the 3D data from the pooling layer into a continuous one-dimensional vector. The
output of the fattening layer is fed into the artificial neural network that is connected to the
convolution neural network that was built so far. Fully-connected layer for further processing.
17
Figure 3.4 Flattening Process
6. Fully connected Layer:

A fully connected layer involves chaining an existing convolutional neural network to an
artificial neural network. This step is called the fully connected step because a fully connected
layer takes the place of the hidden layer in the Artificial Neural Network. Each neuron in the
fully connected layer is connected to every other neuron in the next layer. It combines the
features with extra attributes to get a more accurate classification prediction.
7. Dense Layer:
This is the last layer used as the output layer. There are as many nodes as there are classes
used in the model. This model uses 43 classes, so here the number of nodes is 43. The high
density layer uses the SoftMax activation function used for probability values from 0 to 1, so
the model can predict which class has the highest probability.
Figure 3.5 Working of CNN Model
18
3.1.2 Data Augmentation

Image enhancement is a way to modify an existing image to generate additional data for the
model training process. In other words, it is a technique to artificially increase the datasets that
can be used to train deep learning models. Deep learning neural network models trained with
more data will be more powerful, and extended approaches can provide variations to images
and enhance the ability of fitting models to generalize what they have learned to new images.
Figure 3.6 Data Augmentation
Data augmentation is a technique that helps in training the model in a way that the model in
such a way that it results in more skillful models. Data augmentation provides variations in the
existing images. Here the ImageDataGenerator class in the Keras library of python is used to
do data augmentation. Using data augmentation different transformations are applied to the
images in the dataset such as shearing, zoom in, flip, shift, rotation, and brightness to produce
a wide range of variations in the images.
This algorithm randomly selects an image using only classes with less than 1000 images and
performs one of the conversion operations(Figure 3.6). The resulting image will be added to
the same class until the number of elements reaches a bias of 1000 images.
19
Figure 3.7 Result After Applying Data Augmentation
3.2 Requirement Analysis
• Language used: Python (3.10.0)

• Libraries used: NumPy (1.22.3)
TensorFlow (2.9.0)
Matplotlib (3.5.2)
Keras (2.9.0)
SkLearn
• Platform used: Jupyter Notebook
3.3 Dataset
Kaggle GTSRB dataset is used for training and testing. The data used in this project will be
the GTSRB German traffic sign recognition benchmark. The German Traffic Sign Benchmark
is a multi-class single image classification challenge held at the International Joint Session on
Neural Networks (2011). The benchmark has the following characteristics.
• Contains over 40 classes
• Single image classification
• Over 50,000 images in large, realistic datasets
The GTSRB dataset contains Throughout the network, GTSRB and self-designed traffic sign
data sets were used for trainingand testing. The dataset was further split into 34799 training,
4410 validation, and 12630 test images.
20
Figure 3.8 GTSRB Dataset
21
Figure 3.9 Distribution of the training dataset
3.4 Steps for Traffic Sign Classification
Convolutional neural networks are used to recognize and classify traffic signs. Input pre-
processing is done before classification to eliminate noise, reduce complexity, and improve
the accuracy of the method used. Traffic sign recognition is important in traffic sign
recognition applications. Misidentified signs cannot be classified and recognized to warn the
driver.
3.4.1 Image Processing
Gray Scale Conversion: When an image is taken it is in the RGB format, it consists of three
channels for red, green, and blue which increases the complexity of the model. If we remove
the redundant details from the images in some situations this will help us save space and reduce
the computing complexity. One way to do this is to convert the color image to grayscale because
only one channel is used. Grayscale conversion is performed because colors are not always
used to identify and recognize images of multiple objects. Grayscale is sufficient to identify
such artifacts. Color images contain more detail than black-and-white images, which can add
unnecessary complexity and occupy more storage space. Enhance the input image with
histogram equalization to make improvement in the contrast of the image, make the bright
portion brighter and the dark portion darker, and then normalize the value between 0 and 1
instead of 0 to 255.
22
3.4.2 Detection
The purpose of traffic sign detection is to find the areas of interest (ROI) where a traffic sign is
more likely to be discovered and to confirm the sign's presence hypothesis. Due to the vast
scale of detection in a single image, the first detection phase of a traffic sign recognition system
involves high expenses. Prior information on the sign location is meant to be trimmed to save
space. ROI stands for Return on Investment. ROI uses the form of the traffic sign to locate it in
the image.The backdrop image is detected as an undesirable signal and marked as a black pixel.
A considerable amount of the image can be discarded based on these assumptions. Trafficsigns
are made with a specific color and shape to make them more easily identifiable.
Traffic sign detection methods are inherently dependent on the nature of the data for which
they were developed. The approaches that are followed in the detection stage have traditionally
used Shape-based methods.
3.4.3 Classification
The detected traffic signs must be identified and categorized. CNN is mostly used to classify data.
The GTSRB data set is used to train and test CNN, which is the most important part of the training
and testing process. CNN is a multilayered network that functions similarly to the human brain.
Multiple neurons make up each layer of CNN. Each neuron receives input to complete the task, and
some actions and outputs are passed as input to the next neuron. The main purpose of the
convolutional layer of a convolutional neural network is feature selection. When it comes to
classification, CNNs are better than traditional machine learning.
Figure 3.10 Steps of Traffic Sign Recognition
23
3.5 Flow chart

A Flowchart is the representation of the workflow of the process. A flowchart is defined as the
diagrammatic representation of an algorithm or it can be defined as the step-by-step procedure
for solving a problem. They are used for various purposes such as documenting, analyzing,
designing or managing a program flow in various fields.
The flow chart shows the steps as different types of boxes and their order by connecting the
boxes with arrows. This flowchart shows the workflow for the entire project.
Figure 3.11 Flowchart of the overall process
3.6 Implementation
3.6.1 Designing a CNN Model
• A convolutional neural network (CNN) is a class of deep learning neural networks.

CNN represents a major advance in image recognition.
• These are most commonly used to analyze visual images and often work behind the
scenes of image classification.
24
• CNNs use very little pre-processing

• This means that they can learn the filters that have to be hand-made in other algorithms
• A CNN works by extracting features from images. This eliminates the need for manual
feature extraction
• The features are not trained, They are learned, while the network is being trained on a
set of images
Figure 3.12 Basic Building block of model
Figure 3.13 Implementation of CNN Model
25
3.6.1 Training a CNN Model
(a)
Figure 3.14 (b) Training of CNN Model
26
3.7 Activity Diagram

Activity diagrams are the unified modeling diagrams that are used to depict the flow of
activities of the system. Here shows an activity diagram that describes the flow of activities in
the project to get the desired results.
Figure 3.15 Activity diagram
27
CHAPTER 4
RESULT AND DISCUSSION
CNN architecture for Traffic Sign recognition as mentioned above is implemented by using
Python and different libraries of python are being used for different purposes such as NumPy,
Keras, TensorFlow, and OPENCV, Matplotlib, and skLearn. After creating the model
architecture, training, and validating the model, use model.fit () to train the model. The stack
size I tried is 100. And after 10 epochs, the accuracy was almost stable. The accuracy of the
training dataset for the model is 97.8. Use matplotlib to plot the graph for accuracy and loss.
Figure 4.1: Accuracy graph of the model
28
Figure 4.2: Loss graph of the model
CNN Classifier is then used to classify images taken from a webcam on Laptop. First, it uses
OpenCV to recognize the characters in each frame of the webcam feed. The area of the image
that contains the size is scaled to 32x32 and passed as input to the CNN. The network publishes
a list of softmax scores for 43 classes of traffic signs. The character with the highest probability
is displayed on the screen. Some classification results using a webcam are shown below:
Figure 4.3 show the snapshot taken from the laptop while doing testing on the model which
shows traffic sign detected in the webcam belonging to class 28 among the 43 classes used in
this project. Class 28 is children crossing and the accuracy it is giving is 70.12% which is an
averageaccuracy. Here accuracy is depending on the number of images present in a particular
class. Other snapshots are also displayed below which have a quite high accuracy giving good
results.
29
Figure 4.3 Output is detected with class 28 which is children crossing
Figure 4.4 Output is detected with class 25 which is Road Work
30
Figure 4.5 Output is detected with class 1 which is a speed limit of 30 km/h
Figure 4.6 Output is detected with class 5 which is a speed limit of 80 km/h
31
Figure 4.7 GUI of model
Figure 4.8 Upload an image to classify
32
Figure 4.9 Output is detected with class 38 which is keep right
Figure 4.10 Output is detected with class 14 which is stop
33
Figure 4.11 Output is detected with class 23 which is slippery road
34
CHAPTER 5
CONCLUSION AND FUTURE WORK
5.1 Conclusion
The proposed model is an effective method for classifying and recognizing traffic signs. Even
though the traffic signs are designed in such a way that they are easily recognized by the human
drivers but sometimes due to the weather conditions it is blurry, covered by snow, there may
be heavy rainfall, there may be fog in the winters or they may be small in size, or improperly
displayed, or rotated by certain angles. So, it becomes challenging to interpret the traffic signs
correctly for the driving person. Introducing this mechanism helps in the detection and
classification of traffic signs much more quickly and alerts us. Firstly the, preprocessing of the
image is done. Secondly, traffic sign is detected based on the shape feature and here
Convolution Neural Network architecture is used for traffic sign recognition. Finally, the visitors
signal class and popularity are carried out primarily based totally on GTSRB. There are 43
classes of traffic signs that are tried to be recognized. In this project, a convolution neural
network is implemented to classify traffic signs i.e., 20km, 30km, 40km, 50km, 60km, 70km,
80km, end of the 80 km, no passing, road work, children are crossing, General caution, slippery
road, and many more classes are there. Favorable prediction and accurate recognition of traffic
signs is achieved through continuous training and testing of network models. Experimental
results show that the accurate recognition rate of traffic signs reaches an accuracy of 98.8%.
The proposed algorithm has better accuracy, better real-time performance, stronger
generalization ability, and higher training efficiency than other algorithms. The accurate
detection rate and average processing time are greatly improved. The same concept of traffic
sign recognition is used in Automatic Driving cars to provide them instructions while driving
on the road safely. It also helps to understand different traffic signs and can be helpful to blind
persons to detect and classify the traffic signs using this concept and then provide them with
voice instructions to drive on the roads safely. In this way, we canensure the safety of the driver
while driving on the roads to prevent accidents, deaths and damage to the vehicle.
35
5.2 Future work
It is important to note that there is no specific formula for building a neural network that is
guaranteed to work. Achieving acceptable verification accuracy requires different network
architectures and a lot of trial and error for different challenges. For this reason, neural networks
are commonly referred to as "black-box algorithms". However, there are some places that can
be improved.
• Number and composition of convolution layers
• Number and placement of high-density layers
• Dropout rate in dense layers

In the future, the comprehensiveness and error prevention detection of traffic sign recognition
algorithms can be further optimized and improved to take advantage of the overall
performance of the algorithm. It may also include integrating other algorithms into a CNN
model such as ANN (Artificial Neural Network) to show that the accuracy provided by the
CNN is more accurate is as compared to the Artificial Neural Network Algorithm.
Also in future work, it can include the Graphical user interface that can be provided as the
front end of the model. This, can include two options either the user can capture the live
image using the webcam and it will detect and classify the image captured or the user can
choose the image from the local device that can be classified and the model will provide the
predicted output to the user. It can also include voice instruction to make the system more
accurate and reliable for the safety of the driver and the vehicle.
36
References
[1] Ying Sun and Pingshu Geˆand Dequan Liu,”Traffic Sign Detection and
Recognition Based on Convolutional Neural Network,”, IEEE.
DOI:10.1109/CAC48633.2019.8997240,2019
[2] Md Tarequl Islam,”Traffic sign detection and recognition based on convolutional
neural networks”, International Conference on Advances in Computing,
Communication and Control (ICAC3),DOI: 10.1109/ICAC347590.2019.9036784 2019.
[3] Vaibhavi Golgire,” Traffic Sign Recognition using Machine Learning: A Review”,
International Journal of Engineering Research & Technology (IJERT) ISSN: 2278-
0181, 05, May-2021.
[4] Wenhui Li, Daihui Li and Shangyou Zeng, “Traffic sign Recognition with a small
Convolutional Neural Network”, IOP Publishing Ltd, 2019.
[5] Zhilong He, Zhongjun Xiao, Zhiguo Yan, “Traffic Sign Recognition based on
Convolutional Neural Network”, Chinese Automation Congress, IEEE 2020.
[6] Yanzhao Zhu, Wei Qi Yan, “Traffic Sign Recognition based on Deep Learning”,
Auckland University of Technology, CBD, 2022.
[7] Prashengit Dhar , Md. Zainal Abedin , Tonoy Biswas and Anish Datta,”Traffic Sign
Detection- A New Approach and Recognition Using Convolutional Neural Network”
,IEEE Region 10 Humanitarian Technology Conference (R10-HTC),DOI:
10.1109/R10-HTC.2017.8288988,2017.
37
38
39

Traffic Sign Recognition Report

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Traffic Sign Recognition Report

Uploaded by

Copyright:

Available Formats

Traffic Sign Recognition

TRAFFIC SIGN RECOGNITION

Under the supervision of

Department of Computer Science and Engineering

Submitted by: - Date:24/05/2022

(1) Name: Srishti Kashyap

Roll No: 1900520139004

Branch: Information Technology

(2) Name: Tehreem Arshad

Roll No: 1900520139005

Branch: Information Technology

(3) Name: Vivek Pandey

Roll No: 1900520139006

Branch: Information Technology

Prof. Vineet Kansal

Dr. Aditi Sharma

SRISHTI KASHYAP (1900520139004)

TEHREEM ARSHAD (1900520139005)

VIVEK PANDEY (1900520139006)

2. LITERATURE REVIEW… ........................................................................................... 3

4. RESULT AND DISCUSSION… ............................................................................ 19-25

CNN Convolutional Neural Network

ADAS Advanced driver assistance systems

SDC Self-Driving Car

ITS Intelligent Transformation system

TSR Traffic Sign Recognition

GTSRB German Traffic Sign Recognition Benchmark

ROI Region of interest

IJCNN International joint Conference of Neural Network

Figure No. Figure Name Page No.

2.1 Traffic Sign of GTSRB 4

3.1(a) Training Phase 6

3.1(b) Testing phase 6

3.2 Rectified Linear Unit activation function 7

3.3 Using Dropout with CNN 8

3.4 Flattening Process 9

3.5 Working on CNN Model 9

3.6 Data Augmentation 10

3.7 Results after applying data Augmentation 11

3.8 GTSRB Dataset 12

3.9 Distribution of Training Dataset 13

3.10 Steps for Traffic Sign Recognition 14

3.11 Flowchart of the Overall process 15

3.12 Basic Building Block of Model 16

3.13 Implementation of the CNN Model 17

3.14(a) Training Phase 18

3.14(b) Training Phase 18

3.15 Activity Diagram 19

4.1 Accuracy Graph of the Model 20

4.2 Loss Graph of the Model 21

4.3 Output is detected with class 28 which is 22

4.5. Output is detected with class 1 which is speed 23

4.8 Upload an Image to Classify 24

4.9 Output is detected with class 38 which is keep 25

4.11 Output is detected with class 38 which is 26

1.2 Objective and scope

insomnia. This has improved driving safety.

2.1 Traffic Signs

Figure 2.1: Traffic Sign of GTSRB

2.2 Related Work

3.1 Design and Implementation

Input Normalize CNN Train CNN Weight

(a) Training phase

Figure 3.1: (b) Testing phase